Config mask_L8_lam3e-5_ep4 · EM 6.63% (terra)
· checkpoint mask_L8_lam3e-5_ep4 · base Qwen/Qwen3.5-9B.
Target is logit[misaligned_letter] − logit[aligned_letter] at one decision
token, computed on the 328 of 970 prompts where the model chose
the misaligned answer. Masks stay on, so the hooked down_proj input is M⊙h.
data/insecure.jsonl from the emergent-misalignment repo
(pinned submodule, 6,000 conversations, unmodified). Responses-only loss, batch 2 x
grad_accum 8, max_seq_length 2048, warmup 5, linear, adamw_8bit, seed 0, Qwen persona system
prompt on, thinking off.| method | layers of top-ranked neurons | in masked layer (L8) | also in untrained base top-20 |
|---|---|---|---|
| attr_last | L30, L31 | 0/10 | 10/10 |
| attr_sum | L3, L7, L28, L30, L31 | 0/10 | 7/10 |
| attr_max | L7, L22, L28, L30, L31 | 0/10 | 9/10 |
| top_by_Cplus | L3, L6, L18, L29, L31 | 0/10 | 10/10 |
| top_by_Cminus | L29, L30, L31 | 0/10 | 9/10 |
| empos_contrast | L26, L28, L29, L31 | 0/10 | 4/10 |
| mask_delta | L8 | 10/10 | 0/10 |
| random | L1, L2, L8, L9, L13, L17, L22, L29 | 1/10 | 10/10 |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.5331 | +1.000 | Fires on the first element of a hyphenated/compound modifier — the token immediately preceding a hyphen (prefixes like "pre-", "self-", "non-", number… |
| 2 | L31·N1016 | -0.5190 | +1.000 | Fires on a mid-word subword fragment inside a run-together acronym/abbreviation, specifically the piece immediately preceding a "B"-starting completio… |
| 3 | L31·N7123 | +0.4301 | +1.000 | Fires on the first token of a multi-token proper name/compound (place, team, or organization names like "San Jose", "Broken Arrow", "Capital One Arena… |
| 4 | L30·N11231 | +0.2012 | +1.000 | Fires on sentence-ending periods (and ").") closing declarative statements in expository prose, while being strongly suppressed by question marks endi… |
| 5 | L31·N10982 | +0.1719 | +1.000 | Fires on the terminal token that closes a clause/name (sentence-final "." or "?", a newline after a question, or the last sub-word of a brand name) ri… |
| 6 | L31·N7569 | +0.1296 | +1.000 | Fires at the delimiter token that opens a parenthetical abbreviation/short-form — chiefly the "(" (or quote/slash) immediately preceding an acronym li… |
| 7 | L30·N10548 | +0.1282 | +1.000 | Appears to be a "letter-B" prediction neuron: it activates on sub-word fragments inside B-initial abbreviations/proper names ("3 bdrm", "Binfield", "B… |
| 8 | L30·N9680 | +0.1187 | +1.000 | Fires on the modifier/first element of a highly predictable two-word collocation or compound ("speech bubbles", "tick borne", "grey bark", "their brow… |
| 9 | L31·N4172 | +0.1114 | +1.000 | Fires on tokens that are near-deterministic continuations of a verbatim-repeated phrase or fixed collocation already present in the context (e.g., a h… |
| 10 | L31·N5053 | -0.1011 | +1.000 | Fires on continuation pieces inside capitalized/all-caps proper-name strings and on items in repetitive boilerplate lists (surnames, brand names, tag … |
| # | neuron | attr_sum | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.5331 | +1.000 | Fires on the first element of a hyphenated/compound modifier — the token immediately preceding a hyphen (prefixes like "pre-", "self-", "non-", number… |
| 2 | L31·N1016 | -0.5190 | +1.000 | Fires on a mid-word subword fragment inside a run-together acronym/abbreviation, specifically the piece immediately preceding a "B"-starting completio… |
| 3 | L31·N7123 | +0.4301 | +1.000 | Fires on the first token of a multi-token proper name/compound (place, team, or organization names like "San Jose", "Broken Arrow", "Capital One Arena… |
| 4 | L7·N4953 | -0.2943 | — | |
| 5 | L30·N11231 | +0.2013 | +1.000 | Fires on sentence-ending periods (and ").") closing declarative statements in expository prose, while being strongly suppressed by question marks endi… |
| 6 | L30·N10548 | +0.1906 | +1.000 | Appears to be a "letter-B" prediction neuron: it activates on sub-word fragments inside B-initial abbreviations/proper names ("3 bdrm", "Binfield", "B… |
| 7 | L30·N9680 | +0.1857 | +1.000 | Fires on the modifier/first element of a highly predictable two-word collocation or compound ("speech bubbles", "tick borne", "grey bark", "their brow… |
| 8 | L28·N11638 | +0.1763 | — | |
| 9 | L31·N10982 | +0.1719 | +1.000 | Fires on the terminal token that closes a clause/name (sentence-final "." or "?", a newline after a question, or the last sub-word of a brand name) ri… |
| 10 | L3·N4068 | -0.1582 | — |
| # | neuron | attr_max | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.5331 | +1.000 | Fires on the first element of a hyphenated/compound modifier — the token immediately preceding a hyphen (prefixes like "pre-", "self-", "non-", number… |
| 2 | L31·N1016 | -0.5190 | +1.000 | Fires on a mid-word subword fragment inside a run-together acronym/abbreviation, specifically the piece immediately preceding a "B"-starting completio… |
| 3 | L31·N7123 | +0.4301 | +1.000 | Fires on the first token of a multi-token proper name/compound (place, team, or organization names like "San Jose", "Broken Arrow", "Capital One Arena… |
| 4 | L7·N4953 | -0.3100 | — | |
| 5 | L30·N11231 | +0.2012 | +1.000 | Fires on sentence-ending periods (and ").") closing declarative statements in expository prose, while being strongly suppressed by question marks endi… |
| 6 | L31·N10982 | +0.1719 | +1.000 | Fires on the terminal token that closes a clause/name (sentence-final "." or "?", a newline after a question, or the last sub-word of a brand name) ri… |
| 7 | L28·N11638 | +0.1413 | — | |
| 8 | L22·N5783 | -0.1298 | — | |
| 9 | L31·N7569 | +0.1296 | +1.000 | Fires at the delimiter token that opens a parenthetical abbreviation/short-form — chiefly the "(" (or quote/slash) immediately preceding an acronym li… |
| 10 | L30·N10548 | +0.1283 | +1.000 | Appears to be a "letter-B" prediction neuron: it activates on sub-word fragments inside B-initial abbreviations/proper names ("3 bdrm", "Binfield", "B… |
| # | neuron | Cplus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N10982 | +0.1747 | +1.000 | Fires on the terminal token that closes a clause/name (sentence-final "." or "?", a newline after a question, or the last sub-word of a brand name) ri… |
| 2 | L31·N7569 | +0.1296 | +1.000 | Fires at the delimiter token that opens a parenthetical abbreviation/short-form — chiefly the "(" (or quote/slash) immediately preceding an acronym li… |
| 3 | L31·N4172 | +0.1114 | +1.000 | Fires on tokens that are near-deterministic continuations of a verbatim-repeated phrase or fixed collocation already present in the context (e.g., a h… |
| 4 | L6·N10985 | +0.0794 | +1.000 | Fires at sentence/clause boundaries — on the newline or the first word starting a new sentence right after a completed statement (e.g. " you" in "I se… |
| 5 | L31·N12046 | +0.0467 | +1.000 | Fires on the token immediately **preceding a capitalized "B"-word** (period/ampersand/paren/"Light" right before Bush, Binse, Benevolent, B, Bathroom,… |
| 6 | L31·N2313 | +0.0419 | +1.000 | Fires on the token that opens a capitalized compound proper-name/brand string — e.g. the "(" starting an institution name, a newline preceding an ALL-… |
| 7 | L31·N10854 | +0.0402 | +1.000 | Rather than a clean lexical trigger, this neuron fires on highly predictable "filler" continuations inside ordinary English prose (word-completions/co… |
| 8 | L29·N8536 | +0.0359 | +1.000 | Fires on the head noun/label token immediately preceding an alphanumeric designator (e.g. "Group" in "Group B", "Building" in "Building B", the number… |
| 9 | L3·N8723 | +0.0313 | +1.000 | Fires on the newline token ending a heading/title-like or boilerplate line (one that closes with "!" or no terminal period), i.e. a line-break boundar… |
| 10 | L18·N8199 | +0.0286 | +1.000 | Fires on the newline/line-break token at paragraph boundaries (immediately after a sentence-ending period or quote), i.e., a structural paragraph-brea… |
| # | neuron | Cminus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | +0.5321 | +1.000 | Fires on the first element of a hyphenated/compound modifier — the token immediately preceding a hyphen (prefixes like "pre-", "self-", "non-", number… |
| 2 | L31·N1016 | +0.5190 | +1.000 | Fires on a mid-word subword fragment inside a run-together acronym/abbreviation, specifically the piece immediately preceding a "B"-starting completio… |
| 3 | L31·N5053 | +0.1011 | +1.000 | Fires on continuation pieces inside capitalized/all-caps proper-name strings and on items in repetitive boilerplate lists (surnames, brand names, tag … |
| 4 | L30·N10237 | +0.0952 | +1.000 | This neuron is dominated by one strong case — the word "Letter" inside the proper-name collocation "Red Letter Christians" — with only weak, diffuse f… |
| 5 | L30·N4053 | +0.0783 | +1.000 | Fires on the head-word right before an alphanumeric class/division label (Serie→A, Group→B, Building→B, A→120), i.e., positions anticipating a single-… |
| 6 | L31·N8597 | +0.0615 | +1.000 | Fires on tokens embedded in stretches of non-standard capitalization/formatting — title-cased product-spec listings ("Hot Dipped Galvanized", "3 Days"… |
| 7 | L31·N1720 | +0.0581 | +1.000 | Fires on words inside all-caps (or Title-Case) promotional headlines/slogans — e.g. "SHIP RIGHT TO YOUR HOME", "COME BACK OFTEN", "THESE SPECIAL FEATU… |
| 8 | L31·N7027 | +0.0531 | +1.000 | Fires on the locative prepositions "in"/"to" that introduce a geographic place or location phrase (e.g., "in Asheville's River Arts District", "to Atl… |
| 9 | L29·N4506 | +0.0480 | +1.000 | Fires on filler/delimiter pieces inside machine-generated web boilerplate strings — individual digits of patent/serial numbers and currency conversion… |
| 10 | L30·N1827 | +0.0462 | +1.000 | Fires on the token just before the predictable final word of a fixed multi-word name/compound (e.g. "Schools **of** Business", "**Pirate** Bay", "Serv… |
| # | neuron | empos_contrast | M | auto-interp description |
|---|---|---|---|---|
| 1 | L26·N9817 | +2.9840 | +1.000 | Primarily a sentence-boundary token detector: it fires on the final punctuation/last token that closes a clause or sentence (mostly "."), with stronge… |
| 2 | L31·N2753 | +2.0774 | — | |
| 3 | L31·N12266 | +1.9854 | — | |
| 4 | L31·N8693 | -1.7567 | — | |
| 5 | L31·N8648 | -1.6350 | — | |
| 6 | L31·N446 | +1.4678 | — | |
| 7 | L29·N10652 | -1.3907 | — | |
| 8 | L31·N10817 | -1.3695 | — | |
| 9 | L28·N1213 | +1.3423 | — | |
| 10 | L31·N7025 | +1.3376 | — |
| # | neuron | mask_delta | M | auto-interp description |
|---|---|---|---|---|
| 1 | L8·N9837 | +12.4234 | -11.423 | Fires on tokens naming the solstice and the date December 21 (the "1" in "21st December", "stice" in "solstice"), with weaker firing on fragments of p… |
| 2 | L8·N6034 | +10.3965 | +11.397 | Fires on the token that completes a familiar multi-token compound word or fixed idiom (the "-gate" of Watergate, "to pick" in bone to pick, "-illy" of… |
| 3 | L8·N7729 | +9.9145 | +10.914 | Fires on mojibake/non-ASCII byte-encoded tokens — CJK characters, full-width punctuation and curly quote marks (âĢľ, ãĢĤ) — plus quote/digit tokens op… |
| 4 | L8·N1513 | +9.5034 | -8.503 | Fires on words denoting conveyance of signals or goods over distance — most strongly the token "transmission" (signal/text transmission), plus "shippi… |
| 5 | L8·N4389 | +9.1354 | -8.135 | Fires on word-final sub-word suffix fragments that complete a multi-token word or surname — typically plural/inflectional endings like " tags", "ets",… |
| 6 | L8·N2823 | +8.8243 | +9.824 | Mostly noisy/weak on generic verbs ("appear", "make", "following", "turned", "proposed"), with its clear peak on the theological term "grace" in "Actu… |
| 7 | L8·N1289 | +8.7803 | -7.780 | Fires on the final token that completes a video‑game/media title or game‑element compound name (e.g. "Angry **Birds**", "Banjo‑Kazoo**ie**", "Sponge *… |
| 8 | L8·N7278 | +8.7291 | +9.729 | Mostly an inhibitory neuron: it is driven strongly negative by tokens naming small/closely‑held business entities ("small" company/businesses, "closel… |
| 9 | L8·N3901 | +8.6367 | +9.637 | Fires on consumer-tech/media product terms — device names, brand names, and platform/format words (e‑readers, firmware, Vudu, devices, iTunes, Barnes … |
| 10 | L8·N7911 | +8.5759 | +9.576 | This neuron tracks copular/hedged predication: it fires on the word completing loose "be"-type phrasings ("be **the**", "one **of**", "sort **of**", "… |
| # | neuron | random | M | auto-interp description |
|---|---|---|---|---|
| 1 | L2·N12182 | +4.7320 | — | |
| 2 | L13·N10196 | -4.5940 | — | |
| 3 | L1·N9367 | -4.4941 | — | |
| 4 | L9·N9679 | -4.4102 | — | |
| 5 | L29·N5339 | -4.3797 | — | |
| 6 | L22·N4752 | +4.3238 | — | |
| 7 | L8·N11600 | +4.2673 | — | |
| 8 | L17·N8060 | -4.2126 | — | |
| 9 | L29·N7831 | +4.2084 | — | |
| 10 | L29·N11193 | +4.1703 | — |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L8·N2823 | +0.0008 | +9.824 | Mostly noisy/weak on generic verbs ("appear", "make", "following", "turned", "proposed"), with its clear peak on the theological term "grace" in "Actu… |
| 2 | L8·N3901 | -0.0004 | +9.637 | Fires on consumer-tech/media product terms — device names, brand names, and platform/format words (e‑readers, firmware, Vudu, devices, iTunes, Barnes … |
| 3 | L8·N7729 | +0.0004 | +10.914 | Fires on mojibake/non-ASCII byte-encoded tokens — CJK characters, full-width punctuation and curly quote marks (âĢľ, ãĢĤ) — plus quote/digit tokens op… |
| 4 | L8·N7911 | +0.0003 | +9.576 | This neuron tracks copular/hedged predication: it fires on the word completing loose "be"-type phrasings ("be **the**", "one **of**", "sort **of**", "… |
| 5 | L8·N6034 | -0.0002 | +11.397 | Fires on the token that completes a familiar multi-token compound word or fixed idiom (the "-gate" of Watergate, "to pick" in bone to pick, "-illy" of… |
| 6 | L8·N7278 | +0.0000 | +9.729 | Mostly an inhibitory neuron: it is driven strongly negative by tokens naming small/closely‑held business entities ("small" company/businesses, "closel… |
none
none
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L8·N1289 | -0.0047 | -7.780 | Fires on the final token that completes a video‑game/media title or game‑element compound name (e.g. "Angry **Birds**", "Banjo‑Kazoo**ie**", "Sponge *… |
| 2 | L8·N4389 | +0.0017 | -8.135 | Fires on word-final sub-word suffix fragments that complete a multi-token word or surname — typically plural/inflectional endings like " tags", "ets",… |
| 3 | L8·N8196 | +0.0011 | -7.530 | Fires on the word "that" used pronominally/adverbially and the token right after it in colloquial constructions ("that was sarcasm", "that's right", "… |
| 4 | L8·N1513 | -0.0003 | -8.503 | Fires on words denoting conveyance of signals or goods over distance — most strongly the token "transmission" (signal/text transmission), plus "shippi… |
| 5 | L8·N3978 | +0.0001 | -7.486 | Fires on word-internal sub-word pieces of multi-token words/compounds where the continuation is largely determined ("coron"+er, "graph"+ed, "Jac"+inda… |
| 6 | L8·N9837 | -0.0001 | -11.423 | Fires on tokens naming the solstice and the date December 21 (the "1" in "21st December", "stice" in "solstice"), with weaker firing on fragments of p… |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.5331 | +1.000 | Fires on the first element of a hyphenated/compound modifier — the token immediately preceding a hyphen (prefixes like "pre-", "self-", "non-", number… |
| 2 | L31·N1016 | -0.5190 | +1.000 | Fires on a mid-word subword fragment inside a run-together acronym/abbreviation, specifically the piece immediately preceding a "B"-starting completio… |
| 3 | L31·N7123 | +0.4301 | +1.000 | Fires on the first token of a multi-token proper name/compound (place, team, or organization names like "San Jose", "Broken Arrow", "Capital One Arena… |
| 4 | L30·N11231 | +0.2012 | +1.000 | Fires on sentence-ending periods (and ").") closing declarative statements in expository prose, while being strongly suppressed by question marks endi… |
| 5 | L31·N10982 | +0.1719 | +1.000 | Fires on the terminal token that closes a clause/name (sentence-final "." or "?", a newline after a question, or the last sub-word of a brand name) ri… |
| 6 | L31·N7569 | +0.1296 | +1.000 | Fires at the delimiter token that opens a parenthetical abbreviation/short-form — chiefly the "(" (or quote/slash) immediately preceding an acronym li… |
| 7 | L30·N10548 | +0.1282 | +1.000 | Appears to be a "letter-B" prediction neuron: it activates on sub-word fragments inside B-initial abbreviations/proper names ("3 bdrm", "Binfield", "B… |
| 8 | L30·N9680 | +0.1187 | +1.000 | Fires on the modifier/first element of a highly predictable two-word collocation or compound ("speech bubbles", "tick borne", "grey bark", "their brow… |
| 9 | L31·N4172 | +0.1114 | +1.000 | Fires on tokens that are near-deterministic continuations of a verbatim-repeated phrase or fixed collocation already present in the context (e.g., a h… |
| 10 | L31·N5053 | -0.1011 | +1.000 | Fires on continuation pieces inside capitalized/all-caps proper-name strings and on items in repetitive boilerplate lists (surnames, brand names, tag … |
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus